def _assert_in_default_state(t): t.assertIs(distribution_strategy_context._get_default_replica_context(), distribution_strategy_context.get_replica_context()) t.assertIs(None, distribution_strategy_context.get_cross_replica_context()) t.assertIs(distribution_strategy_context._get_default_distribution_strategy(), distribution_strategy_context.get_distribution_strategy()) t.assertFalse(distribution_strategy_context.has_distribution_strategy())
def merge_fn(dist, s): self.assertIs( distribution_strategy_context._get_default_distribution_strategy(), dist) self.assertIs(None, distribution_strategy_context.get_replica_context()) self.assertIs(dist, distribution_strategy_context.get_cross_replica_context()) self.assertIs(dist, distribution_strategy_context.get_distribution_strategy()) self.assertFalse( distribution_strategy_context.has_distribution_strategy()) return "foo_" + s
def run_fn(): replica_context = distribution_strategy_context.get_replica_context() self.assertTrue(replica_context is not None) self.assertIs(None, distribution_strategy_context.get_cross_replica_context()) self.assertTrue(distribution_strategy_context.has_distribution_strategy()) self.assertIs(dist, distribution_strategy_context.get_distribution_strategy()) self.assertEqual("foo", replica_context.merge_call(None, test_arg="foo")) expected_value = _get_test_variable( "bar", variable_scope.VariableSynchronization.AUTO, variable_scope.VariableAggregation.NONE) self.assertDictEqual(expected_value, variable_scope.variable(1.0, name="bar"))
def testScope(self): _assert_in_default_state(self) dist = _TestStrategy() with dist.scope(): self.assertIs(None, distribution_strategy_context.get_replica_context()) self.assertIs(dist, distribution_strategy_context.get_cross_replica_context()) self.assertTrue(distribution_strategy_context.has_distribution_strategy()) self.assertIs(dist, distribution_strategy_context.get_distribution_strategy()) expected_value = _get_test_variable( "baz", variable_scope.VariableSynchronization.AUTO, variable_scope.VariableAggregation.NONE) self.assertDictEqual(expected_value, variable_scope.variable(1.0, name="baz")) _assert_in_default_state(self)
def apply_gradients(self, grads_and_vars, global_step=None, name=None): """Apply gradients to variables. This is the second part of `minimize()`. It returns an `Operation` that applies gradients. Args: grads_and_vars: List of (gradient, variable) pairs as returned by `compute_gradients()`. global_step: Optional `Variable` to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the `Optimizer` constructor. Returns: An `Operation` that applies the specified gradients. If `global_step` was not None, that operation also increments `global_step`. Raises: TypeError: If `grads_and_vars` is malformed. ValueError: If none of the variables have gradients. RuntimeError: If you should use `_distributed_apply()` instead. """ # This is a default implementation of apply_gradients() that can be shared # by most optimizers. It relies on the subclass implementing the following # methods: _create_slots(), _prepare(), _apply_dense(), and _apply_sparse(). # Handle DistributionStrategy case. if distribute_ctx.get_cross_replica_context(): raise RuntimeError("Use `_distributed_apply()` instead of " "`apply_gradients()` in a cross-replica context.") # TODO(isaprykin): Get rid of `has_distribution_strategy()` check by # always calling _distributed_apply(), using the default distribution # as needed. if distribute_ctx.has_distribution_strategy(): grads_and_vars = get_filtered_grad_fn(lambda: grads_and_vars)() return distribute_ctx.get_replica_context().merge_call( self._distributed_apply, args=(grads_and_vars, global_step, name)) # No DistributionStrategy case. grads_and_vars = tuple(grads_and_vars) # Make sure repeat iteration works. if not grads_and_vars: raise ValueError("No variables provided.") converted_grads_and_vars = [] for g, v in grads_and_vars: if g is not None: try: # Convert the grad to Tensor or IndexedSlices if necessary. g = ops.convert_to_tensor_or_indexed_slices(g) except TypeError: raise TypeError( "Gradient must be convertible to a Tensor" " or IndexedSlices, or None: %s" % g) if not isinstance(g, (ops.Tensor, ops.IndexedSlices)): raise TypeError( "Gradient must be a Tensor, IndexedSlices, or None: %s" % g) p = _get_processor(v) converted_grads_and_vars.append((g, v, p)) converted_grads_and_vars = tuple(converted_grads_and_vars) var_list = [v for g, v, _ in converted_grads_and_vars if g is not None] if not var_list: raise ValueError("No gradients provided for any variable: %s." % ([str(v) for _, v, _ in converted_grads_and_vars],)) with ops.init_scope(): self._create_slots(var_list) update_ops = [] with ops.name_scope(name, self._name) as name: self._prepare() for grad, var, processor in converted_grads_and_vars: if grad is None: continue # We colocate all ops created in _apply_dense or _apply_sparse # on the same device as the variable. # TODO(apassos): figure out how to get the variable name here. if context.executing_eagerly() or isinstance( var, resource_variable_ops.ResourceVariable) and not var._in_graph_mode: # pylint: disable=protected-access scope_name = "" else: scope_name = var.op.name with ops.name_scope("update_" + scope_name), ops.colocate_with(var): update_ops.append(processor.update_op(self, grad)) if global_step is None: apply_updates = self._finish(update_ops, name) else: with ops.control_dependencies([self._finish(update_ops, "update")]): with ops.colocate_with(global_step): if isinstance(global_step, resource_variable_ops.ResourceVariable): # TODO(apassos): the implicit read in assign_add is slow; consider # making it less so. apply_updates = resource_variable_ops.assign_add_variable_op( global_step.handle, ops.convert_to_tensor(1, dtype=global_step.dtype), name=name) else: apply_updates = state_ops.assign_add(global_step, 1, name=name) if not context.executing_eagerly(): if isinstance(apply_updates, ops.Tensor): apply_updates = apply_updates.op train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP) if apply_updates not in train_op: train_op.append(apply_updates) return apply_updates
def init_from_checkpoint(ckpt_dir_or_file, assignment_map): """Initializes current variables with tensors loaded from given checkpoint. Note: This overrides default initialization ops of specified variables and redefines dtype. Assignment map supports following syntax: * `'checkpoint_scope_name/': 'scope_name/'` - will load all variables in current `scope_name` from `checkpoint_scope_name` with matching tensor names. * `'checkpoint_scope_name/some_other_variable': 'scope_name/variable_name'` - will initialize `scope_name/variable_name` variable from `checkpoint_scope_name/some_other_variable`. * `'scope_variable_name': variable` - will initialize given `tf.Variable` object with tensor 'scope_variable_name' from the checkpoint. * `'scope_variable_name': list(variable)` - will initialize list of partitioned variables with tensor 'scope_variable_name' from the checkpoint. * `'/': 'scope_name/'` - will load all variables in current `scope_name` from checkpoint's root (e.g. no scope). Supports loading into partitioned variables, which are represented as `'<variable>/part_<part #>'`. Example: ```python # Say, '/tmp/model.ckpt' has the following tensors: # -- name='old_scope_1/var1', shape=[20, 2] # -- name='old_scope_1/var2', shape=[50, 4] # -- name='old_scope_2/var3', shape=[100, 100] # Create new model's variables with tf.variable_scope('new_scope_1'): var1 = tf.get_variable('var1', shape=[20, 2], initializer=tf.zeros_initializer()) with tf.variable_scope('new_scope_2'): var2 = tf.get_variable('var2', shape=[50, 4], initializer=tf.zeros_initializer()) # Partition into 5 variables along the first axis. var3 = tf.get_variable(name='var3', shape=[100, 100], initializer=tf.zeros_initializer(), partitioner=lambda shape, dtype: [5, 1]) # Initialize all variables in `new_scope_1` from `old_scope_1`. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/': 'new_scope_1'}) # Use names to specify which variables to initialize from checkpoint. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/var1': 'new_scope_1/var1', 'old_scope_1/var2': 'new_scope_2/var2'}) # Or use tf.Variable objects to identify what to initialize. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_1/var1': var1, 'old_scope_1/var2': var2}) # Initialize partitioned variables using variable's name init_from_checkpoint('/tmp/model.ckpt', {'old_scope_2/var3': 'new_scope_2/var3'}) # Or specify the list of tf.Variable objects. init_from_checkpoint('/tmp/model.ckpt', {'old_scope_2/var3': var3._get_variable_list()}) ``` Args: ckpt_dir_or_file: Directory with checkpoints file or path to checkpoint. assignment_map: Dict, where keys are names of the variables in the checkpoint and values are current variables or names of current variables (in default graph). Raises: tf.errors.OpError: If missing checkpoints or tensors in checkpoints. ValueError: If missing variables in current graph. """ if distribution_strategy_context.get_cross_replica_context(): _init_from_checkpoint(None, ckpt_dir_or_file, assignment_map) else: distribution_strategy_context.get_replica_context().merge_call( _init_from_checkpoint, ckpt_dir_or_file, assignment_map)
def assign_moving_average(variable, value, decay, zero_debias=True, name=None): """Compute the moving average of a variable. The moving average of 'variable' updated with 'value' is: variable * decay + value * (1 - decay) The returned Operation sets 'variable' to the newly computed moving average, by performing this subtraction: variable -= (1 - decay) * (variable - value) Since variables that are initialized to a `0` value will be `0` biased, `zero_debias` optionally enables scaling by the mathematically correct debiasing factor of 1 - decay ** num_updates See `ADAM: A Method for Stochastic Optimization` Section 3 for more details (https://arxiv.org/abs/1412.6980). The names of the debias shadow variables, by default, include both the scope they were created in and the scope of the variables they debias. They are also given a uniquifying-suffix. E.g.: ``` with tf.variable_scope('scope1'): with tf.variable_scope('scope2'): var = tf.get_variable('foo') update_1 = tf.assign_moving_average(var, 0.0, 1.0) update_2 = tf.assign_moving_average(var, 0.0, 0.9) # var.name: 'scope1/scope2/foo' # shadow var names: 'scope1/scope2/scope1/scope2/foo/biased' # 'scope1/scope2/scope1/scope2/foo/biased_1' ``` Args: variable: A Variable. value: A tensor with the same shape as 'variable'. decay: A float Tensor or float value. The moving average decay. zero_debias: A python bool. If true, assume the variable is 0-initialized and unbias it, as in https://arxiv.org/abs/1412.6980. See docstring in `_zero_debias` for more details. name: Optional name of the returned operation. Returns: A tensor which if evaluated will compute and return the new moving average. """ def update_fn(v, value, decay=decay): decay = ops.convert_to_tensor(1.0 - decay, name="decay") if decay.dtype != v.dtype.base_dtype: decay = math_ops.cast(decay, v.dtype.base_dtype) if zero_debias: update_delta = _zero_debias(v, value, decay) else: update_delta = (v - value) * decay return state_ops.assign_sub(v, update_delta, name=scope) with ops.name_scope(name, "AssignMovingAvg", [variable, value, decay]) as scope: replica_context = distribution_strategy_context.get_replica_context() if replica_context: # In a replica context, we update variable using the mean of value across # replicas. def merge_fn(strategy, v, value): value = strategy.extended.reduce_to( ds_reduce_util.ReduceOp.MEAN, value, v) return strategy.update(v, update_fn, value) return replica_context.merge_call(merge_fn, args=(variable, value)) else: strategy = distribution_strategy_context.get_cross_replica_context() return strategy.update(variable, update_fn, value)